MSL: BDA function parameters should probably be emitted as pointers instead of references

[Hugobros3]

I've ran into a number of issues with MoltenVK having to do with the way it emits pointers in function parameters, as C++ references instead of C-style pointers. For example this snippet:

%my_kernel_1 = OpFunction %void None %31
         %32 = OpFunctionParameter %_ptr_Function__struct_3
          %a = OpFunctionParameter %int
          %b = OpFunctionParameter %_ptr_PhysicalStorageBuffer_int
%my_kernel_2 = OpLabel
         %38 = OpAccessChain %_ptr_Function_uint %32 %uint_0
         %39 = OpLoad %uint %38
         %43 = OpConvertPtrToU %ulong %b
         %44 = OpExtInst %void %40 1 %41 %a %43
               OpStore %38 %39
               OpReturn
               OpFunctionEnd

gets translated as:

static inline __attribute__((always_inline))
void my_kernel(thread _3& _32, int a, device const int& b)
{
    debugPrintfEXT("hi %d 0x%lx\n", a, reinterpret_cast<ulong>(b));
    _32._m0 = _32._m0;
}

This leads to the following bugs:

• References to void are forbidden but OpTypePointer ... ... %void is legal SPIR-V that SPIRV-Cross can't translate
• Broken OpPtrToU: That reinterpret_cast above fails because b is a reference and can't decay to a pointer. You'd need to take the address of b
• Broken ptr arithmetic for the same reason

Here's a godbolt with a few of these problems condensed to a single sample: https://godbolt.org/z/Tanvao3aT

I'm unsure why this choice was made, perhaps it was a restriction in older Metal ?

[Hugobros3]

This might be what #2368 was running into, but that's only a hunch